Machine learning triumphs in differentiating fungal and bacterial infections.

Introduction

Multiple myeloma (MM), a plasma cell cancer, is increasing in prevalence due to an aging population and improved survival rates. While advances in treatment have raised the 5-year relative survival rate to over 50% [ 1 ], infections remain a major cause of morbidity and mortality in MM patients [ 2 ]. Bacterial infections are particularly common, but recent evidence highlights a growing concern: invasive fungal infections (IFIs).

Historically, studies reported low rates of invasive fungal disease (IFD) in MM patients [ 3 – 5 ]. However, recent data show a notable rise in IFI prevalence, from < 1% [ 6 ] to 3.4% [ 7 ], with invasive pulmonary aspergillosis occurring in 6.2% of cases [ 8 ]. The respiratory tract and bloodstream are the most frequent sites of infection, each accounting for 40% of IFD cases [ 7 ]. Among heavily treated patients (≥ 3 lines of therapy), IFD incidence can reach 15% [ 4 ]. These infections pose severe health risks, contributing to higher mortality in hematologic malignancies [ 6 ]. Specifically, MM patients with IFIs face a 44% all-cause mortality rate within 30 days [ 4 ].

Diagnostic challenges further complicate IFD management. Early-stage symptoms are often nonspecific, overlapping with bacterial infections, while serum galactomannan testing is prone to false positives, leading to unnecessary interventions [ 9 ]. The lack of rapid diagnostic tools forces reliance on low-sensitivity fungal cultures and empirical evidence, increasing the risk of: unnecessary antifungal use (promoting resistance), inappropriate antibiotic prescriptions, and delayed effective treatment, which correlates with higher mortality [ 10 ].

Thus, developing early, accurate predictive models to differentiate IFD from bacterial infections in MM patients could enable timely targeted therapy, rational antibiotic/antifungal use, and improved patient outcomes, reducing mortality and enhancing clinical decision-making.

Polymerase chain reaction tests have demonstrated high sensitivity in early IFI detection, yet the absence of methodological standardization causes lab-to-lab result discrepancies. Next-generation sequencing is capable of simultaneously detecting multiple pathogens from different taxa in a single sample, but it is expensive and has struggled to gain wider acceptance [ 11 ]. There are other new methods for rapid differentiation of fungal and bacterial infections, but large-scale clinical validation has not yet been fully developed [ 12 , 13 ]. Therefore, in recent years, more and more research has focused on biomarkers that can assist in the early detection of IFI and distinguish it from bacterial infections [ 14 , 15 ]. C-reactive protein (CRP)and procalcitonin are used in tandem for the early recognition of fungal and bacterial infections in immunocompromised patients [ 16 , 17 ]. The small sample size, however, restricts its broader use. Earlier research has utilized imaging indicators to differentiate bacterial pneumonia from IFD [ 18 , 19 ], yet there are no distinct and highly specific imaging indicators in the initial stages of IFD, making the applicability of this model debatable. While traditional linear prediction models are mediocre in terms of judgment power, machine learning models show significant advantages in handling large data sets and complex variable relationships [ 20 ]. It has been noted that machine learning can enhance disease prediction models’ accuracy [ 21 ].

Using machine learning models to predict infections in MM patients is becoming more common. Some studies have utilized machine learning algorithms to develop infection prediction models specifically for MM patients [ 22 ]. However, these models usually predict the risk of infection without distinguishing the type of infection. Efforts in the past have been centered on the development and validation of computational models for predicting bacteremia and fungemia using routine electronic medical record data available at the time of ordering blood cultures [ 23 ]. However, blood cultures are time-consuming, have low positivity rates, and high contamination rates. In addition, the predictive model, which was designed primarily for the general hospitalized population, is too broad and not suitable for immunocompromised patients, such as those with MM. Due to the limited sample sizes in previous studies, the relationship between MM and IFD has been examined only through subgroup analyses or by identifying risk factors for IFD in MM patients [ 4 , 7 , 8 , 24 ]. Few studies have applied machine learning techniques to anticipate bacterial and IFIs in this patient population.

The aim of this study was to construct a machine-learning model that leverages multidimensional data, comprising biomarkers and clinical features, for the early recognition of IFD and bacterial infections in MM patients. Through a retrospective cohort analysis, we systematically evaluated a set of critical variables capable of differentiating between IFD and bacterial infections, subsequently prioritizing these variables based on their importance. Various machine learning algorithms were employed to construct the classification model, with the optimal model selected to enhance discrimination accuracy. The resulting predictive model facilitates the early identification of high-risk patients for IFD, guiding the precise administration of antifungal agents and promoting the judicious management of antibiotics. It is anticipated that this application will optimize infection management strategies for MM patients, thereby mitigating the risks associated with antimicrobial drug misuse and reducing patient mortality.

Materials and methods

Patient selection and criteria for inclusion and exclusion

This retrospective case-control study was conducted at the Department of Hematology of the First Affiliated Hospital of Chongqing Medical University, a 3,200-bed tertiary care institution. The study included all inpatients aged 18 years and older diagnosed with MM between October 2012 and October 2023, according to the International Myeloma Working Group (IMWG) criteria [ 25 ]. The exclusion criteria encompass the following: patients with incomplete medical records; individuals lacking essential research indicators, such as MM classification, infection-related pathogenicity or inflammatory markers, and follow-up outcomes; patients who exhibit no clinical evidence of infection despite having positive pathogen cultures indicative of specimen contamination; and patients diagnosed with active malignancies other than MM. Each hospitalization was treated as an independent event, and re-hospitalizations with additional treatment were recorded separately.

The case group (IFD cohort, n = 140) consisted of all consecutive eligible MM patients diagnosed with IFD during the study period. IFD was strictly defined and classified as either proven or probable according to the 2019 European Organization for Research and Treatment of Cancer/Mycoses Study Group (EORTC/MSG) guidelines [ 26 ]; cases classified as possible IFD were excluded. The diagnosis date was determined as the point when clinical, radiological, or microbiological evidence first indicated IFD. Given the rarity and severity of IFD, all eligible cases were included to maximize statistical power and representativeness.

The control group (bacterial infection cohort, n = 158) was selected from MM patients with bacterial infections during the same period at the same institution. Bacterial infections were confirmed by microbiological documentation (e.g., positive cultures from sterile sites) or radiological evidence (e.g., pneumonia or sinusitis) that necessitated antibiotic treatment [ 27 ]. A detailed breakdown of specific infection diagnoses for both the IFD case and bacterial control cohorts is provided in Supplementary Table 1 . Although initial plans aimed for 1:1 matching, the control group was intentionally expanded to improve statistical robustness and better reflect the real-world clinical scenario where bacterial infections are more prevalent. The final case-control ratio of approximately 1:1.13 is well within accepted standards for such studies. All controls satisfied the same inclusion and exclusion criteria as the cases, except for the type of infection, ensuring that both groups were derived from the same source population and remained comparable.

Data collection

This study categorized therapeutic regimens for MM based on International Myeloma Working Group guidelines [ 28 ] into four types: monoclonal antibody protocols (alone or with other agents), combined IMiD and PI therapies (optionally with dexamethasone), IMiD-centered treatments (as monotherapy or with dexamethasone), and PI-focused regimens. The investigation utilized the therapeutic line definition from the International Myeloma Working Group for clinical assessments, including disease progression, intervention analysis, and recurrence monitoring.

All predictive factors were established prior to the study outcome, not during hospitalization. The infection diagnosis date was when the clinician first noted the infection based on symptoms, signs, and microbiological evidence. Key factors included demographic characteristics (age, gender), baseline comorbidities (e.g., diabetes, heart failure, chronic lung disease), and tumor staging (e.g., R-ISS, ISS, Durie-Salmon, cytogenetic abnormalities), assessed at admission or the latest evaluation before infection. Treatment history encompassed antibiotic use in the month prior, treatment plans during the infection (including cycles and lines), stem cell transplantation history, glucocorticoid doses in the 30 and 60 days before infection, and recent surgeries (within 2 weeks), all recorded before the infection. Laboratory indicators (e.g., myeloma cell ratio, β2 microglobulin, neutrophil count, creatinine, free light chain ratio, platelet count, immunoglobulin, hemoglobin, serum albumin, lactate dehydrogenase, blood calcium, C-reactive protein, procalcitonin, complement, liver and kidney function, myocardial enzymes) were based on the latest test results within 7 days prior to infection. Invasive procedures (e.g., central venous catheterization, arterial catheterization, mechanical ventilation, urinary catheterization) and organ support (e.g., total parenteral nutrition, renal replacement therapy, ICU admission, iron overload) were noted if performed before the infection. Disease status (e.g., recurrence, progression, extramedullary infiltration, treatment-related adverse events) and complications (e.g., neutropenia, hypogammaglobulinemia) were assessed based on the most recent evaluation prior to infection. The Eastern Cooperative Oncology Group Performance Status (ECOG) score was also recorded at admission or before the infection.

Overview of ML models

During the data preprocessing phase, a comprehensive multi-level imputation strategy was employed. Features exhibiting more than 40% missing values were excluded to uphold data quality (refer to Supplementary Fig. 1 ). For the remaining variables, the imputation approach was tailored based on the type of variable and the extent of missingness: continuous variables were imputed using the median for those with ≤ 5% missing data, K-Nearest Neighbors (KNN) for those with 5% < missing data ≤ 20%, and Multiple Imputation by Chained Equations (MICE) for those with more than 20% missing data. Categorical variables were imputed using the mode. Post-imputation, integer variables underwent rounding to ensure consistency. This methodology resulted in a fully imputed dataset that maintained the original distribution, thereby facilitating high-quality training for machine learning models. For feature selection, a multi-stage screening strategy was implemented to develop a robust and interpretable feature set. Initially, Pearson correlation coefficients were utilized to filter raw features, retaining those significantly associated with the target variable (|r| > 0.1, p < 0.05) to reduce complexity and eliminate irrelevant variables.

To mitigate potential selection bias associated with the use of a single algorithm, this study employs a triad of complementary feature selection methods for ensemble evaluation: the Boruta algorithm, which utilizes random forests to discern genuinely relevant variables by comparing the importance of original features against “shadow features”; Lasso regression, which leverages L1 regularization to automatically identify key variables through sparse coefficients; and a Logistic Regression (LR) model employing recursive feature elimination (RFE). To enhance robustness and interpretability, a consensus strategy was implemented, whereby only features selected by at least two of the methods were retained to constitute the final feature set. This ensemble approach effectively reduced dimensionality while preserving predictive performance, thereby enhancing model stability and generalization capability.

Subsequently, a binary prediction model was developed using selected clinical features to differentiate IFD from bacterial infections in MM patients. The study, implemented in Python 3.9.15, used stratified random sampling (7:3 ratio) to create training and test sets, maintaining consistent case distribution. Nine classification algorithms were compared: LR, random forest, gradient-boosted trees, XGBoost, LightGBM, SVM with RBF kernel, KNN, AdaBoost, and decision trees. Models underwent 5-fold stratified cross-validation on the training set, with performance assessed on the test set using accuracy, precision, recall, specificity, F1 score, and area under the receiver operating characteristic curve (AUC). Finally, after selecting the best-performing model, the Shapley Additive exPlanations (SHAP) method was used to interpret its predictive behavior, offering insights into feature importance. Bootstrap resampling ( n = 500) quantified uncertainty, producing 95% confidence intervals to evaluate model stability. A detailed visualization report was generated, featuring ROC curves, performance heatmaps, and feature importance plots.

Statistical analysis

Patient data were classified into continuous and categorical variables. The normality of data distribution was evaluated using the Kolmogorov-Smirnov test. Continuous variables following a normal distribution were presented as mean ± standard deviation and analyzed via the t-test. In contrast, the Mann–Whitney U test was employed for continuous variables that did not conform to a normal distribution, with results expressed as median (interquartile range). Categorical data were described as counts and frequencies, and comparisons were conducted using the Chi-square test or Fisher’s exact test. All statistical analyses were two-tailed, with P-values less than 0.05 considered statistically significant. Statistical analyses were performed using Version 4.5.2 of the R statistical software (R Foundation). The ensemble model was developed using Python (Version 3.9.15, Python Software Foundation, Wilmington, DE, USA, https://www.python.org ).

Results

The study screened MM patients, identifying 170 with IFD and 885 with bacterial infections. After excluding 206 cases—62 for incomplete records, 120 for pathogen cultures without clinical infection evidence, and 24 for other active malignancies—140 fungal infection cases were included as the case group. From the bacterial infection group, 158 cases were matched based on concurrent hospitalization and the absence of antifungal medication use, forming the control group. Ultimately, 298 cases (140 fungal and 158 bacterial) were analyzed. Figure 1 outlines the screening process, and baseline characteristics are compared in Supplementary Table 2 .

Patients with IFD had more advanced disease, showing higher relapse rates (16.43% vs. 5.70%), and were more often at ISS stage III (63.57% vs. 48.10%) and Durie-Salmon stage III (88.57% vs. 79.11%). They had a higher prevalence of IgG-type myeloma (64.29% vs. 41.77%) and more frequently exhibited progressive/refractory disease (34.29% vs. 12.03%) and extramedullary infiltration (35.00% vs. 17.09%). Clinically, IFD patients had worse performance status, more comorbidities like hypertension (45.00% vs. 25.32%), renal impairment (37.14% vs. 20.89%), and heart failure (27.86% vs. 11.39%), and a higher incidence of endotoxic shock (16.43% vs. 0.63%).

The IFD cohort experienced significantly more intensive myeloma-directed therapy. They were more likely to receive monoclonal antibody regimens (9.29% vs. 2.53%) and less likely to have no chemotherapy (14.29% vs. 32.91%). Corticosteroid use was higher, with prolonged use (62.14% vs. 48.10%) and greater cumulative doses over 30 days (400.00 vs. 253.00 mg) and 60 days (583.00 vs. 299.50 mg). Stem cell transplantation was more common in the IFD group (20.72% vs. 3.80%), with both autologous (16.43% vs. 3.80%) and allogeneic (4.29% vs. 0%) transplants, and recent transplants within 30 days were more frequent (10.71% vs. 1.90%).

Patients with IFD had significantly higher prior exposure to antibiotics and intensive supportive care compared to others. Antibiotic use was notably more common in the IFD group, with 93.57% having a history of antibiotic therapy in the past month, including long-term combination use (14.29% vs. 0%). IFD patients also required more intensive interventions such as ICU admission (15.00% vs. 0%), urinary catheters (20.71% vs. 3.80%), and central venous catheters (35.71% vs. 12.66%). Additionally, they had higher rates of invasive mechanical ventilation, total parenteral nutrition, and oxygen therapy. A history of fungal infection was also more prevalent in the IFD group (26.43% vs. 1.90%).

Laboratory analyses indicated that patients with IFD demonstrated significant systemic dysregulation across several domains. Hematological parameters were notably compromised, evidenced by reduced levels of hemoglobin, platelets, white blood cells, and absolute lymphocytes, in addition to prolonged neutropenia. Nutritional status was also adversely affected, as indicated by diminished serum albumin, a lower albumin-globulin ratio, and reduced corrected calcium levels. Furthermore, markers of inflammation and organ injury were elevated, including lactate dehydrogenase, total and direct bilirubin, international normalized ratio (INR), and B-type natriuretic peptide (BNP). Additionally, a distinct immunoglobulin profile was identified, characterized by decreased IgA, increased IgG, and a lower prevalence of hypogammaglobulinemia.

IFD patients had a history of more intensive healthcare use, with higher median hospitalizations (4.50 vs. 2.00), more treatment cycles (4.00 vs. 1.00), more lines of prior myeloma therapies (2.00 vs. 1.00), and longer hospital stays (24.50 vs. 15.00 days).

Evaluation of the associations between diverse biomarkers

To mitigate the complexity and dimensionality of subsequent analyses, this study initially conducted a preliminary screening of all raw features utilizing Pearson’s correlation coefficient. Linear correlations between the 93 features and the target variable were calculated, with selection criteria grounded in correlation strength and statistical significance. Specifically, features exhibiting a significant correlation with the target variable were retained, defined by an absolute correlation coefficient of |r| > 0.1 and a p-value < 0.05. This procedure successfully identified 50 key features for further in-depth analysis. Figure 2 (A) displays the Feature Significance Analysis Plot, while Fig. 2 (B) shows the feature correlation heatmap.

Variables ranked by significance

To address the limitations inherent in a single methodological approach, we employed a concurrent application of three distinct machine learning algorithms for comprehensive screening: the Boruta algorithm, Lasso regression, and recursive feature elimination. These methodologies identified 19, 22, and 10 significant features, respectively (Fig. 3 A and D). By implementing an ensemble consensus strategy—retaining only those features recognized by at least two of the algorithms—we ultimately identified 16 key variables characterized by high stability and interpretability for the construction of the final predictive model. The identified features included: prior antibiotic therapy in previous month, ECOG performance status, duration of hospitalisation, prior fungal infection, BNP, corrected serum calcium, use of mask or nasal oxygen inhalation, serum albumin, number of lines of myeloma therapies, number of hospitalizations in the past two years, lymphocyte percentage, hemoglobin level, long-term and combined use of multiple antibiotics, immunoglobulin A, albumin-globulin ratio, and immunoglobulin G (Fig. 3 E).

The machine learning model’s prediction outcome

The dataset was divided into a training set and an independent test set using stratified random sampling at a 7:3 ratio to preserve the original distribution of IFD versus bacterial infection cases. Prior to model development, all features were standardized using z-score normalization. This study employed and compared nine machine learning algorithms: LR, Support Vector Machine (SVM), KNN, Random Forest (RF), Gradient Boosting, LightGBM, XGBoost, AdaBoost, and Decision Tree. Each model was trained on the standardized feature data and evaluated for generalization using stratified 5-fold cross-validation (Stratified 5-fold CV), with final performance assessed on an independent test set comprising 30% of the total data. The comprehensive results are presented in Table 1 . The AUC values for the nine models ranged from 0.860 to 0.967, with the LR model demonstrating superior performance. Figure 4 A illustrates the ROC curves for each model on the test set, annotated with their respective cross-validation AUC averages. Figure 4 B presents the precision-recall trade-off plot. Notably, the LR model achieves the highest recall (Recall = 0.908) among all candidate models, positioned at the far right of the graph, indicating the lowest rate of missed positive cases. Concurrently, its precision (Precision = 0.919) remains high, comparable to other high-performing models such as XGBoost and SVM. Moreover, the LR model attains an AUC value of 0.967, the highest among all models, further substantiating its robust overall discrimination capability.

Stability verification of the optimal model

To assess the stability of the optimal LR model’s performance, a stratified bootstrap sampling technique was utilized, involving 500 resampling evaluations on the test dataset. The aggregated results of the bootstrap sampling for key performance metrics are detailed in Table 2 , including the mean, 95% confidence interval (CI), and standard deviation (SD). The results are as follows: the AUC was 0.951 (95% CI: 0.940–0.961, SD: 0.005), accuracy was 0.907 (95% CI: 0.867–0.933, SD: 0.018), precision was 0.904 (95% CI: 0.862–0.929, SD: 0.019), recall was 0.898 (95% CI: 0.810–0.952, SD: 0.039), the F1 score was 0.900 (95% CI: 0.852–0.930, SD: 0.021), and specificity was 0.916 (95% CI: 0.875–0.938, SD: 0.019). The standard deviation for all metrics remained below 0.04, with overall narrow 95% confidence intervals. Notably, the AUC confidence interval width was only 0.021, suggesting minimal performance variation across different data subsets and high estimation precision. Figure 5 illustrates that the cumulative mean of the AUC converged after approximately 50 iterations during the bootstrap sampling analysis. Subsequently, the 95% confidence interval narrowed and stabilized, indicating that the resampling process had reached a state of convergence.

Feature importance analysis (SHAP)

The interpretability of machine learning models is essential for their clinical application. To improve the interpretability of the most effective LR model, SHAP analysis was employed to quantify the relative contribution of each predictor to the model’s output. The mean absolute SHAP values were used to identify the 16 most influential features (Fig. 6 A). These features, ranked by their importance, are as follows: prior antibiotic therapy in the preceding month, ECOG performance status, duration of hospitalization, corrected serum calcium levels, number of hospitalizations over two years, lymphocyte percentage, serum albumin levels, hemoglobin levels, history of fungal infection, number of lines of myeloma therapies, immunoglobulin A levels, immunoglobulin G levels, albumin-globulin ratio, BNP levels, long-term and combined use of multiple antibiotics, and the use of a mask or nasal oxygen inhalation. The SHAP summary plot (Fig. 6 B) provides a visual representation of the magnitude and direction of each feature’s contribution to the model output across all samples. Notable features contributing significantly include prior antibiotic therapy within the preceding month, elevated ECOG performance status, extended duration of hospitalization, increased frequency of hospitalizations over two years, history of fungal infections, increased lines of myeloma therapies, elevated immunoglobulin G and BNP levels, prolonged and combined use of multiple antibiotics, and the use of masks or nasal oxygen inhalation, all of which were associated with a heightened predictive probability of invasive fungal infection in patients with MM. Conversely, higher levels of corrected serum calcium, lymphocyte percentage, serum albumin, hemoglobin, immunoglobulin A, and albumin-globulin ratio were linked to an increased predictive probability of bacterial infection in these patients. The SHAP decision plot (Fig. 6 C) further elucidates consistent feature contribution trajectories across patients, with prior antibiotic therapy in the preceding month, corrected serum calcium, and ECOG performance status frequently emerging as dominant factors in the risk prediction pathways.

Discussion

Patients with MM are susceptible to a spectrum of infections due to impaired immune function. The clinical manifestations of invasive fungal and bacterial infections frequently overlap, yet the therapeutic approaches and prognoses can differ significantly. Immunocompromised patients with IFD experience a notably high mortality rate [ 6 ], underscoring the importance of early diagnosis and prompt antifungal treatment in reducing mortality. Traditional microbiological assays are often time-consuming and lack sufficient sensitivity. In our study, we employed various machine learning algorithms to develop predictive models, assessing their performance through cross-validation and other evaluation techniques. The results demonstrate that the model exhibits high accuracy and stability in distinguishing between the two types of infections. By accurately identifying pathogen types, the model can inform targeted anti-infective therapies and mitigate the risk of drug resistance associated with empirical treatment.

Previous research has successfully developed predictive models to distinguish between bacterial and fungal infections [ 23 ]. However, two primary limitations persist: first, the variables employed in these models often lack clinical information, utilize a single data modality, and do not provide sufficient granularity. Second, there is a notable absence of specific validation for hematological diseases, particularly MM. Moreover, the majority of existing infection risk prediction models for patients with MM are confined to overall risk assessments and do not facilitate precise classification of pathogen types [ 22 , 29 ]. This study is, to our knowledge, the first to create a model aimed at identifying infection types specifically in MM patients, addressing a major gap in existing research.

In this study, we conducted an analysis of a dataset consisting of 140 patients diagnosed with IFD and 158 patients with bacterial infections, incorporating both clinical and laboratory data. We developed and validated nine machine learning models to evaluate the risk of differentiating between these infections. The AUC values for these models ranged from 0.860 to 0.967, suggesting their potential for clinical application. We anticipate that the predictive accuracy of these models will be further enhanced with the integration of additional clinical data and supplementary features. In the domain of predictive modeling, achieving high predictive accuracy is often the primary objective, with the AUC frequently utilized by statisticians as an evaluation metric. However, in clinical contexts, metrics such as sensitivity or positive predictive value may hold greater significance [ 30 ]. In our analysis, the nine models demonstrated accuracy ranging from 0.847 to 0.918, recall between 0.817 and 0.908, and AUC values between 0.860 and 0.967, indicating a robust capability to distinguish bacterial infections from IFD.

A comparative analysis of various models reveals that the LR algorithm exhibits superior performance, as evidenced by its comprehensive metrics: an accuracy of 0.918, precision of 0.919, recall of 0.908, F1 score of 0.913, and an AUC of 0.967. By thoroughly analyzing feature correlations within high-dimensional data, this model effectively addresses the limitations and subjectivity inherent in traditional single-indicator diagnostic methods. It shows particularly notable advantages in identifying cases when atypical symptoms manifest in the early stages of IFD. To enhance the interpretability of the LR model, SHAP analysis was employed to quantify the relative contribution of each predictor to the model’s output.

Feature selection is crucial for model development. We assessed numerous clinical features to find those linked to invasive fungal and bacterial infections, including clinical signs and lab results. Key factors identified were treatment history, immune status, clinical traits, and inflammatory markers. These enhance model accuracy and offer valuable diagnostic guidance for clinicians.

Numerous early investigations have demonstrated that patients diagnosed with IFD underwent a greater number of chemotherapy regimens compared to those with non-IFD [ 7 , 8 ]. The incidence of IFI among patients receiving three or more lines of therapy was reported to be 15.0% [ 4 ]. Furthermore, a significant proportion of patients with IFI experienced multiple hospitalizations within a two-year timeframe [ 31 ]. Extended hospitalization and recurrent admissions heighten the risk of pathogen exposure, thereby increasing the likelihood of infection. In the present study, a comparison between patients with IFD and those with bacterial infections indicated that the former group received a greater number of myeloma therapies and experienced a higher frequency of hospitalizations within the two-year period, a finding that aligns closely with the previously mentioned research outcomes.

Furthermore, the variables incorporated into our predictive model—specifically, recent antibiotic therapy within the past month, prolonged concurrent use of multiple antibiotics, a history of fungal infection, duration of hospitalization, and ECOG performance status—are consistent with factors identified in multicenter prospective studies. In Colombia, extended hospitalization, abdominal surgery, meropenem administration, and hemodialysis have been recognized as significant risk factors for candidemia [ 32 ]. Additionally, the duration of prior antibiotic therapy has been shown to be independently associated with candidemia [ 33 ], and treatment regimens lasting ten days or more have been included in predictive models for candidemia and overall mortality [ 34 ]. Research indicates a notable correlation between higher ECOG performance status and an increased incidence of IFD. Additionally, these studies reveal a significant association between a history of prior fungal infections and an elevated risk of IFD [ 3 ], suggesting that previous exposure to fungal pathogens may have a lasting impact on host susceptibility through mechanisms such as immune memory remodeling, a finding that aligns with our observations. Furthermore, invasive medical procedures can compromise innate immune defense mechanisms. The damage to the nasal mucosa resulting from the use of nasal oxygen tubes may create an environment conducive to fungal invasion. Previous case reports have substantiated this pathway as a contributing factor to the development of IFIs [ 35 ].

Abnormal immunoglobulins are significant predictors of immune deficiency in MM patients. Bacterial infections are often linked to low IgG levels, while IFIs are more complex, potentially due to broader immune deficiencies and interference from the disease itself. Recent studies, including Nakamura et al. [ 36 ], highlight that MM can affect fungal serological diagnoses, leading to false positives in IgG-type patients. Additionally, IgG is crucial for systemic anti-infective immunity, while IgA and IgG antibodies react to commensal fungi, indicating that different immunoglobulin types may serve distinct roles against various pathogens [ 37 ].

Albumin levels and the albumin-globulin ratio (AGR) are essential biomarkers indicative of nutritional status and systemic inflammation. Numerous studies have established a strong association between hypoalbuminemia and poor prognosis [ 31 ]. For example, in patients with mucormycosis, severe hypoalbuminemia has been identified as a significant metabolic abnormality and serves as an independent prognostic marker across three distinct cohorts [ 38 ]. In a study by Tsai CK [ 24 ], factors associated with fungal infections in patients with MM were investigated, revealing that light chain disease, anemia, hypoalbuminemia, and undergoing allogeneic stem cell transplantation were independent predictors of fungal infection. These findings align closely with the results of our study. MM is characterized by the excessive production of monoclonal immunoglobulin (i.e., M protein), resulting in elevated globulin levels, while albumin levels are often reduced due to chronic inflammation or hepatic impairment. Consequently, a decreased AGR is commonly observed among patients with MM.

The observed reduction in lymphocyte percentage within the IFD group of our study is likely attributable to the increased frequency of monoclonal antibody administration in these patients. This hypothesis is supported by Lim’s research, which reports an IFD prevalence of 7% among MM patients undergoing monoclonal antibody treatment regimens [ 7 ]. Additionally, other studies on IFD have corroborated the association with lymphopenia [ 39 ]. Aspergillus infection can lead to severe myocardial injury, which is a manifestation of invasive aspergillosis [ 40 ]. In such cases, BNP, a marker of myocardial injury, may be significantly elevated, indicating the extent of cardiac involvement.

This model notably integrates variables that have been insufficiently explored in prior research, including adjustments for serum calcium levels. Evidence suggests that calcium is not only a vital nutrient for eukaryotic cells but also serves as a crucial second messenger in the regulation of numerous physiological processes [ 41 ]. In human pathogenic fungi, calcium signaling pathways, particularly the Ca²⁺-calcineurin pathway, are pivotal in mediating responses to various host environmental stresses, such as membrane damage, osmotic pressure, and oxidative stress. For instance, Candida albicans utilizes the Ca²⁺-calcineurin pathway to sustain virulence and resist antifungal drug pressure [ 42 ]. Future investigations should further explore the mechanisms of calcium signaling pathways in fungal infections, as this could lead to significant advancements in fundamental research and have profound implications for antifungal therapy and clinical drug development.

The predictive model developed in this study exhibits practical value in differentiating between invasive fungal infections and bacterial infections in patients with multiple myeloma. Nonetheless, the research is subject to several limitations. Firstly, the data were sourced exclusively from a single large academic medical centre, lacking external validation datasets. This limitation obscures the model’s applicability to other healthcare settings and necessitates further validation to assess its generalizability. Second, the retrospective design limited the inclusion of prospective data and prevented a comparative analysis of clinical outcomes between patients managed with and without the guidance of the model, thus restricting the reliability of the evidence. A prospective interventional study comparing users and non-users of the model would yield more robust evidence of its clinical utility. However, conducting such research necessitates thorough ethical approval, a process that is both extensive and time-consuming at our institution. We are currently in the process of preparing the requisite documentation for ethical submission and intend to undertake a prospective cohort study upon receiving approval. Third, owing to data limitations, we were unable to conduct a systematic evaluation of the model’s impact on key clinical process indicators, including diagnostic delay, length of hospital stay, timing of appropriate treatment initiation, and clinical efficacy. These outcomes are crucial for assessing whether the model provides tangible real-world benefits. Fourth, while the LR model showed strong discriminatory power (AUC 0.967) and stable performance with bootstrap sampling (standard deviations < 0.04 and narrow 95% confidence intervals), it lacked formal calibration assessment. Calibration is essential in predictive models, particularly for clinical decision-making, to ensure predicted probabilities match observed outcomes. Without metrics like calibration curves or Brier scores, we cannot verify if the predicted probabilities accurately represent actual event rates, which is crucial given the linear assumptions of LR. Future research should address these limitations by undertaking the following actions: conducting comprehensive model validation using multicenter external datasets to evaluate both discriminative ability and calibration; performing prospective cohort studies that compare clinical outcomes with and without the application of the model, including a systematic evaluation of clinical process indicators such as diagnostic delay, length of stay, and time to appropriate treatment; integrating multimodal data, such as imaging, with machine learning techniques to enhance model performance; and incorporating formal calibration analyses, including calibration curves, Hosmer-Lemeshow tests, and Brier scores, as well as employing calibration techniques like Platt scaling or isotonic regression, to improve model reliability and clinical applicability.

In conclusion, the predictive model developed through machine learning techniques in this study offers a novel approach for distinguishing between invasive fungal and bacterial infections in patients with MM, demonstrating potential for clinical application. Nevertheless, further research is necessary to address its limitations and facilitate broader implementation and dissemination.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1

Acknowledgements

We extend our sincere gratitude to all the patients who participated in this study, as well as to the staff of the Information Department at the First Affiliated Hospital of Chongqing Medical University for their invaluable contributions.

Author contributions

XP and LC contributed equally as co-first authors. WZ and XH contributed equally as Co-corresponding authors. XH, LC, XP, and WZ designed and performed the study. XP collected the data. LC, XP, and XZ are responsible for the processing of statistical data and the development of research algorithms. XP and LC drafted the manuscript under the guidance of XH. All authors contributed to the article and approved the submitted version.

Funding

National Natural Science Foundation of China, Project No. 62376044.

Data availability

The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study adhered to the principles outlined in the Declaration of Helsinki and received approval from the Human Ethics Review Committee of the First Affiliated Hospital of Chongqing Medical University (reference number: K2023–314). The ethics review board granted an exemption from obtaining informed consent due to the retrospective nature of the study. Patient records were collected and analyzed with strict measures to maintain the confidentiality of participants’ identities, thereby ensuring the protection of patient privacy.

Consent for publication

Not applicable.

Artificial intelligence (AI)-assisted technology

In this study, machine learning algorithms (Random Forest, XGBoost, Logistic Regression, Naive Bayes, Gradient Boosting, SVM) were employed to develop a predictive model for distinguishing invasive fungal from bacterial infections in multiple myeloma patients. No LLMs (e.g., ChatGPT) were used for manuscript preparation.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Contributor Information

Weimin Zhu, Email: zwmcyc@sina.com.

Xue Hu, Email: huxue@hospital.cqmu.edu.cn.

References

1. Pulte D, Jansen L, Castro FA, Emrich K, Katalinic A, Holleczek B, Brenner H. Trends in survival of multiple myeloma patients in Germany and the United States in the first decade of the 21st century. Br J Haematol. 2015;171(2):189–96.

2. Blimark C, Holmberg E, Mellqvist UH, Landgren O, Björkholm M, Hultcrantz M, Kjellander C, Turesson I, Kristinsson SY. Multiple myeloma and infections: a population-based study on 9253 multiple myeloma patients. Haematologica. 2015;100(1):107–13.

3. Liu J, Huang H, Li Y, Liu L, Li J, Liu Z, Lu J, Ouyang J, Song Y, Zhou D, et al. Epidemiology and treatment of invasive fungal diseases in patients with multiple myeloma: findings from a multicenter prospective study from China. Tumour Biol. 2016;37(6):7893–900.

4. Teh BW, Teng JC, Urbancic K, Grigg A, Harrison SJ, Worth LJ, Slavin MA, Thursky KA. Invasive fungal infections in patients with multiple myeloma: a multi-center study in the era of novel myeloma therapies. Haematologica. 2015;100(1):e28–31.

5. Tisi MC, Hohaus S, Cuccaro A, Innocenti I, De Carolis E, Za T, D’Alò F, Laurenti L, Fianchi L, Sica S, et al. Invasive fungal infections in chronic lymphoproliferative disorders: a monocentric retrospective study. Haematologica. 2017;102(3):e108–11.

6. Pagano L, Caira M, Candoni A, Offidani M, Fianchi L, Martino B, Pastore D, Picardi M, Bonini A, Chierichini A, et al. The epidemiology of fungal infections in patients with hematologic malignancies: the SEIFEM-2004 study. Haematologica. 2006;91(8):1068–75.

7. Lim C, Sinha P, Harrison SJ, Quach H, Slavin MA, Teh BW. Low rates of invasive fungal disease in patients with multiple myeloma managed with new generation therapies: Results from a multi-centre cohort study. Mycoses. 2021;64(1):30–4.

8. Fuchs E, Lavi N, Carasso Y, Oren I, Hardak E. Invasive pulmonary aspergillosis in multiple myeloma patients: a sizeable diagnosis in the era of novel anti-myeloma therapies. Mycoses. 2021;64(10):1298–303.

9. Duarte RF, Sánchez-Ortega I, Cuesta I, Arnan M, Patiño B, Fernández de Sevilla A, Gudiol C, Ayats J, Cuenca-Estrella M. Serum galactomannan-based early detection of invasive aspergillosis in hematology patients receiving effective antimold prophylaxis. Clin Infect Dis. 2014;59(12):1696–702.

10. Wolf S, Weber S, Janetta A, Klein F, Enssle JC, Hogardt M, Kempf VAJ, Kessel J, Vehreschild M, Steffen B, et al. Epidemiology and outcomes of Candida spp. bloodstream infections in cancer patients: a comparative retrospective study from a German tertiary cancer center. Infection. 2025.

11. Fang W, Wu J, Cheng M, Zhu X, Du M, Chen C, Liao W, Zhi K, Pan W. Diagnosis of invasive fungal infections: challenges and recent developments. J Biomed Sci. 2023;30(1):42.

12. Wang J, Meng S, Lin K, Yi X, Sun Y, Xu X, He N, Zhang Z, Hu H, Qie X, et al. Leveraging single-cell Raman spectroscopy and single-cell sorting for the detection and identification of yeast infections. Anal Chim Acta. 2023;1239:340658.

13. Lee PW, Totten M, Traylor A, Zhang SX, Wang TH, Hsieh K. Cross-kingdom pathogen detection via duplex universal PCR and high-resolution melt. Biosens Bioelectron. 2025;270:116922.

14. Yang YC, Mao J. Value of platelet count in the early diagnosis of nosocomial invasive fungal infections in premature infants. Platelets. 2018;29(1):65–70.

15. Niu D, Huang Q, Yang F, Tian W, Li C, Ding L, et al. Serum biomarkers to differentiate gram-negative, gram-positive and fungal infection in febrile patients. J Med Microbiol. 2021;70(7).

16. Liu Y, Zhang X, Yue T, Tang Y, Ke Z, Li Y, Luo X, Huang L. Combination of C-reactive protein and procalcitonin in distinguishing fungal from bacterial infections early in immunocompromised children. Antibiot (Basel) 2022;11(6).

17. Stoma I, Karpov I, Uss A, Krivenko S, Iskrov I, Milanovich N, Vlasenkova S, Lendina I, Belyavskaya K, Cherniak V. Combination of sepsis biomarkers may indicate an invasive fungal infection in haematological patients. Biomarkers. 2019;24(4):401–6.

18. Gong M, Xu J, Li K, Li K, Xia Y, Jing Y, Chen J, Li J, Yang J, Du M, et al. A CT-based nomogram for differentiating invasive fungal disease of the lung from bacterial pneumonia. BMC Med Imaging. 2022;22(1):172.

19. Wang F, Li X, Wen R, Luo H, Liu D, Qi S, Jing Y, Wang P, Deng G, Huang C, et al. Pneumonia-Plus: a deep learning model for the classification of bacterial, fungal, and viral pneumonia based on CT tomography. Eur Radiol. 2023;33(12):8869–78.

20. Goecks J, Jalili V, Heiser LM, Gray JW. How Machine Learning Will Transform Biomedicine Cell. 2020;181(1):92–101.

21. Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Hardt M, Liu PJ, Liu X, Marcus J, Sun M, et al. Scalable and accurate deep learning with electronic health records. NPJ Digit Med. 2018;1:18.

22. Peng T, Liu L, Liu F, Ding L, Liu J, Zhou H, Liu C. Machine learning-based infection prediction model for newly diagnosed multiple myeloma patients. Front Neuroinform. 2022;16:1063610.

23. Bhavani SV, Lonjers Z, Carey KA, Afshar M, Gilbert ER, Shah NS, Huang ES, Churpek MM. The development and validation of a machine learning model to predict bacteremia and fungemia in hospitalized patients using electronic health record data. Crit Care Med. 2020;48(11):e1020–8.

24. Tsai CK, Liu YC, Kuan AS, Lee KL, Yeh CM, Lee YT, Hsiao LT, Ko PS, Wang HY, Chen PM, et al. Risk and impact of invasive fungal infections in patients with multiple myeloma. Ann Hematol. 2020;99(8):1813–22.

25. Criteria for the classification of. monoclonal gammopathies, multiple myeloma and related disorders: a report of the International Myeloma Working Group. Br J Haematol. 2003;121(5):749–57.

26. Donnelly JP, Chen SC, Kauffman CA, Steinbach WJ, Baddley JW, Verweij PE, Clancy CJ, Wingard JR, Lockhart SR, Groll AH, et al. Revision and update of the consensus definitions of invasive fungal disease from the european organization for research and treatment of cancer and the mycoses study group education and research consortium. Clin Infect Dis. 2020;71(6):1367–76.

27. van der Galiën HT, Loeffen EAH, Miedema KGE, Tissing WJE. Predictive value of PCT and IL-6 for bacterial infection in children with cancer and febrile neutropenia. Support Care Cancer. 2018;26(11):3819–26.

28. Kumar S, Paiva B, Anderson KC, Durie B, Landgren O, Moreau P, Munshi N, Lonial S, Bladé J, Mateos MV, et al. International myeloma working group consensus criteria for response and minimal residual disease assessment in multiple myeloma. Lancet Oncol. 2016;17(8):e328–46.

29. Encinas C, Hernandez-Rivas J, Oriol A, Rosiñol L, Blanchard MJ, Bellón JM, García-Sanz R, de la Rubia J, de la Guía AL, Jímenez-Ubieto A, et al. A simple score to predict early severe infections in patients with newly diagnosed multiple myeloma. Blood Cancer J. 2022;12(4):68.

30. Rose S. Machine Learning for Prediction in Electronic Health Data. JAMA Netw Open. 2018;1(4):e181404.

31. Li Y, Wu Y, Gao Y, Niu X, Li J, Tang M, Fu C, Qi R, Song B, Chen H, et al. Machine-learning based prediction of prognostic risk factors in patients with invasive candidiasis infection and bacterial bloodstream infection: a singled centered retrospective study. BMC Infect Dis. 2022;22(1):150.

32. Ortíz Ruiz G, Osorio J, Valderrama S, Álvarez D, Elías Díaz R, Calderón J, Ballesteros D, Franco A. Risk factors for candidemia in non-neutropenic critical patients in Colombia. Med Intensiva. 2016;40(3):139–44.

33. Dudoignon E, Alanio A, Anstey J, Depret F, Coutrot M, Fratani A, Jully M, Cupaciu A, Chaussard M, Oueslati H, et al. Outcome and potentially modifiable risk factors for candidemia in critically ill burns patients: A matched cohort study. Mycoses. 2019;62(3):237–46.

34. Keighley C, Chen SC, Marriott D, Pope A, Chapman B, Kennedy K, Bak N, Underwood N, Wilson HL, McDonald K, et al. Candidaemia and a risk predictive model for overall mortality: a prospective multicentre study. BMC Infect Dis. 2019;19(1):445.

35. Shaikh BS, Appelbaum PC, Jones JM, Christiansen D. Colonization of nasal ulcers as a source of Candida parapsilosis fungemia. Arch Otolaryngol. 1980;106(7):434–6.

36. Nakamura S, Maeda Y, Sumitani R, Oura M, Sogabe K, Yagi H, Fujii S, Harada T, Matsuoka KI, Miki H. False-positive galactomannan test results in multiple myeloma. Diseases 2025;13(4).

37. Doron I, Kusakabe T, Iliev ID. Immunoglobulins at the interface of the gut mycobiota and anti-fungal immunity. Semin Immunol. 2023;67:101757.

38. Pikoulas A, Morianos I, Nidris V, Hamdy R, Intze E, López-López Á, Moran-Garrido M, Muthu V, Halabalaki M, Papaioanou V, et al. Albumin orchestrates a natural host defence mechanism against mucormycosis. Nature. 2026;649(8097):693–702.

39. Alali M, Giurcanu M, Elmuti L, Kumar M. Pediatric invasive fungal risk score in cancer and hematopoietic stem cell transplantation patients with febrile neutropenia. J Pediatr Hematol Oncol. 2022;44(2):e334–42.

40. Ge Y, Zou L, Xue J, Cai D, Wang Y, Han H, Hu N. Aspergillus-induced cardiac injury modeling and preventive strategy screening via a universal precise cardiomyocyte-based electrophysiological biosensing platform. ACS Sens. 2025;10(10):7366–77.

41. Zhang C, Ren Y, Gu H, Gao L, Zhang Y, Lu L. Calcineurin-mediated intracellular organelle calcium homeostasis is required for the survival of fungal pathogens upon extracellular calcium stimuli. Virulence. 2021;12(1):1091–110.

42. Li W, Shrivastava M, Lu H, Jiang Y. Calcium-calcineurin signaling pathway in Candida albicans: A potential drug target. Microbiol Res. 2021;249:126786.

Associated Data

Supplementary Materials

Supplementary Material 1

Data Availability Statement

The datasets generated and analyzed during the current study are available from the corresponding author on reasonable request.